OpenAI wants agents to move past software engineers and into accounting, medicine, investing, and every other white collar job. The company's own adoption numbers, reported this week, show how far that push still has to travel, including inside OpenAI itself.
The products and the reach
ChatGPT Work launched last month at $20 a month, OpenAI's lowest subscription tier. Paired with Codex, the company's agentic coding tool now available as desktop and mobile apps, the two products have a combined 20 million users. That sits against a much larger base of over a billion people who have prompted ChatGPT in some form.
Twenty million agent users out of a billion plus chat users is the first gap. The second one is sharper.
Internal use versus paying customers
| Codex usage | |
|---|---|
| OpenAI employees, internally, June 2026 | 98% |
| Organizational subscribers | 17% |
| Individual subscribers | under 1% |
OpenAI's own staff have all but fully adopted the tool that is supposed to be the model for how everyone else will work with AI agents. Its paying customers, the people the product is actually built for, have not followed anywhere close to the same curve.
What OpenAI says the plan is
Product lead Thibault Sottiaux framed the push in mission terms: "It's the very mission of OpenAI, to bring everyone along." Andrew Ambrosino, the lead engineer on the desktop app, described the harder design question behind that goal: "We have to always parse out, are we doing the workflow that everybody else will be doing?"
That second quote is the more useful one. It names the actual problem, which is not that agents lack capability. Software engineers already get real value from Codex today. The problem is that most other jobs do not come with a command line, a repository, or a habit of writing structured instructions for a tool to execute, so the same product needs an interface most professions have never had a reason to develop.
Why the internal number is 98 percent and the external one is not
Engineers inside OpenAI are close to the best-case user for an agent tool. They already think in discrete, well-specified tasks, they are comfortable iterating when a first attempt is wrong, and they work inside a codebase and workflow the product was purpose-built around. That combination is rare outside a technical team, and it explains the gap better than any story about the product being unfinished.
This lines up with what a Princeton and UC San Diego study found about AI agent skill libraries more broadly, that structure is what actually determines whether an agent performs well, not raw capability or how much context it is given. OpenAI's own employees bring that structure by habit. A first-time user in accounting or medicine has to be taught it, which is a UX problem before it is a capability one.
What the gap means before you roll out agents company wide
Two things follow for a team evaluating agent adoption rather than reading about OpenAI's.
A 98 percent internal number from the company that built the tool is not evidence the tool works for a general audience, it is evidence the tool works well for people who already think the way the tool expects. Do not treat a vendor's own dogfooding numbers as a preview of your organization's adoption curve unless your team's daily work resembles theirs.
Pilot with the people in your organization closest to that engineer profile first, the ones already comfortable breaking a task into steps and iterating on a wrong first answer, before expanding to a team that has never worked that way. The adoption curve inside OpenAI suggests that gap in working style, not the model's raw ability, is the actual bottleneck to plan around.
The rule worth keeping
A company selling agents to everyone and using them internally at 98 percent is not a contradiction. It is a preview of exactly which users an agent already works for today, and a warning about how much interface and habit-building work stands between that user and everyone else. If you are budgeting an agent rollout, AI Stack Optimizer can help work out which parts of your stack are already earning their keep before you add one more tool to the pile.